ECE 595, Section 10 Numerical Simulations Lecture 2: Problems in Numerical Computing

Prof. Peter Bermel January 9, 2013

Outline
Overall Goals Finding Special Values Fourier Transforms Eigenproblems Ordinary Differential Equations Partial Differential Equations

1/9/2013

ECE 595, Prof. Bermel

Recap: Goals for This Class

Learn/review key mathematics Learn widely-used numerical techniques Become a capable user of this software Appreciate strengths and weaknesses of competing algorithms Convey your research results to an audience of colleagues
1/9/2013 ECE 595, Prof. Bermel

RW Hamming (left), developing errorcorrecting codes (AT&T)

Finding Special Values

Finding zeros
1D versus multidimensional Speed versus certainty

Finding minima and maxima (optimization)

Convex vs. non-convex problems Global vs. local search Derivative vs. non-derivative search

1/9/2013

ECE 595, Prof. Bermel

Finding Zeros
Key concept: bracketing Bisection continuously halve intervals Brents method adds inverse quadratic interpolation Newton-Raphson method uses tangent Laguerres method assume spacing of roots at a and b:

1/9/2013

ECE 595, Prof. Bermel

Finding Minima (or Maxima)

Golden Section Search Brents Method Downhill Simplex Conjugate gradient methods Multiple level, single linkage (MLSL)

These and further images from Numerical Recipes, by WH Press et al.

1/9/2013

ECE 595, Prof. Bermel

Recap: Fourier Transforms

J.W. Cooley (IEEE Global History Network)

1/9/2013

ECE 595, Prof. Bermel

Cooley-Tukey Algorithm

1/9/2013

ECE 595, Prof. Bermel

Recap: Eigenproblems

1/9/2013

ECE 595, Prof. Bermel

Eigenproblems
Direct method: solve det(A-1)=0 Similarity transformations: A Z-1AZ
Atomic transformations: construct each Z explicitly Factorization methods: QR and QL methods

1/9/2013

ECE 595, Prof. Bermel

Atomic Transformations in Eigenproblems

Jacobi

Householder

Keep iterating until off-diagonal elements are small, or use factorization approach
1/9/2013 ECE 595, Prof. Bermel

Factorization in Eigenproblems
Most common approach known as QR method Can also do the same with Slow in general, but fast in certain cases:
Tridiagonal matrices Hessenberg matrix

1/9/2013

ECE 595, Prof. Bermel

Ordinary Differential Equations

Euler method nave rearrangement of ODE Runge-Kutta methods match multiple Euler steps to a higher-order Taylor expansion Richardson extrapolation extrapolate computed value to 0 step size Predictor-corrector methods store solution to extrapolate next point, and then correct it

1/9/2013

ECE 595, Prof. Bermel

ODE Boundary Value Problems

Shooting Method

Boundary Value Method

1/9/2013

ECE 595, Prof. Bermel

Partial Differential Equations

Classes: parabolic, hyperbolic, and elliptic Initial value vs. boundary value problems Finite difference Finite element methods Monte-Carlo Spectral Variational methods
ECE 595, Prof. Bermel

1/9/2013

Next Class
Introduction to computational complexity Please read Chapter 1 of Computational Complexity: A Modern Approach by Arora & Barak

1/9/2013

ECE 595, Prof. Bermel